Residual stream
Every layer reads from the same running total and writes back into it. It is the model's working memory, and most of what you want to know is already written on it in plain sight.
Nothing inside a model is hidden.It is only unlit.
An attempt to make mechanistic interpretability something you can poke at, not only something you read about.
Interpretability has a distribution problem.
The field produces real insight. It arrives as PDFs.
You do not build intuition for a residual stream by reading about a residual stream. You build it the way you built intuition for anything else: by moving something and watching what breaks.
So the bet here is narrow. Take the ideas that already exist, and make them touchable. A slider instead of a screenshot. A live activation instead of figure four.
Where to put your hands.
Four ideas carry most of the weight in this field. None of them originate here.
Attention heads
A head decides which earlier tokens the current one is allowed to look at. Some heads do something crisp enough to name: previous-token, induction, attention sink.
Features
Directions in activation space that stand for something. There are far more things worth representing than dimensions to spare, so directions overlap. That crowding is called superposition.
Circuits
When heads and MLPs compose into a path that produces one specific behaviour, that path is a circuit. Find it, cut it, and watch the behaviour go with it. That last step is the part worth building for.
Built in public, and early.
What this is
A public notebook for learning mechanistic interpretability by building the instruments, instead of only reading the results.
Small reproducible experiments on open-weight models, with the code attached to every claim.
Aimed at one question that people shipping LLM apps actually ask: why did my model do that?
What this is not
A product. There is nothing to buy, no tiered waitlist, no dashboard behind a login.
A research lab. Nothing written here should be cited as a result.
Finished, or even shaped. The right form for this tool is still open, and one possible answer is that it should not exist.
The first experiments are not up yet.
Leave an address and you get one message when they are. Nothing else.